Reproduce “"I am bad": Interpreting Stealthy, Universal and Robust Audio Jailbreaks in Audio-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Isha Gupta, David Khachaturov, Robert Mullins
Researchers demonstrated the first universal audio jailbreaks that bypass safety mechanisms in Audio-Language Models (ALMs), remaining stealthy and robust against real-world degradations. These attacks achieved up to 65% attack success rate and notably revealed that ALMs interpret imperceptible adversarial noise as coherent, toxic first-person speech.
Key insights
- Audio-Language Models interpret imperceptible adversarial audio as coherent, often toxic, first-person linguistic content, effectively hijacking the model's linguistic priors.
- Stealthy audio jailbreaks exhibit universality, generalizing across diverse textual prompts, tasks, and unseen base audio samples.
- These adversarial attacks demonstrate high robustness against common real-world audio degradations, making them practical threats.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.